0 Replies Latest reply: Sep 12, 2013 1:35 PM by ChasSoto RSS

    ZFSSA 7410 vs. slow Macintosh SMB performance


      We ran into an issue only recently because our new building has Cisco edge switches connected to our building router via 10Gbps uplinks (each edge port is 1Gbps).  Macs (10.7.x & 10.8.x) connected this way exhibit very slow (300Kbps!) SMB performance to the 7410 cluster.  When connected directly to the router via 1Gig ports, or to a switch with single- or dual-connected 1Gig uplinks, performance is just fine.  After some network tracing and sleuthing, we determined that it's a result of "delayed ACKs" from the Macs to the ZFSSA.  By setting net.inet.tcp.delayed_ack=0 on the Macs, we get decent performance.  It only happens against our ZFSSA - 10Gig connected Windows and Linux/Samba servers work just fine.


      We are wondering if anyone else has experienced this, and if there was a fix on the ZFSSA that you applied to correct this.  Solaris 11 appears to have a related  tcp_deferred_ack_interval tuning parameter (http://docs.oracle.com/cd/E19455-01/816-0607/6m735r5ga/) that might be the right approach (but naturally the AK is "closed" so I can't test it without getting slapped!).  I have a case open with Oracle, but perhaps you can help us point them in the right direction.